Where the model is less sure.
An age model is only as good as the photos it learned from and the photos it receives. Here are its weak spots, with numbers.
Updated 2026-10-18
What the model learned from
FairFace was trained on 108,501 photos from the Flickr YFCC-100M collection. Its authors built it to balance seven population groups, because older datasets were dominated by light-skinned faces. Balance on paper does not guarantee equal accuracy, so we measured it.
Gaps between population groups
On 2,236 adult faces never used for training, the model alone gave the right group or the next one for 94.5 % to 98.5 % of faces depending on the group. The largest gap is 4.1 points, under the 10-point limit we set before measuring. Between women and men, the full site pipeline found the exact group 56.1 % and 66.3 % of the time: the confidence intervals overlap, but we keep watching it.
Older ages are harder
Accuracy drops with age. The model alone gives the right group or the next one for 98.2 % of faces aged 20-29, but 89.8 % at 50-59 and 91.2 % at 60-69. The 70+ group was measured on only 35 faces: too few to conclude, and it is not measured yet on the full site pipeline.
Faces that are not found
On our test images, no face was found in 7.5 % of cases, above our 5 % target. Most were in profile, turned down or partly hidden. Photos with several people are refused on purpose.
What we have not measured yet
- Real selfies: the test set comes from the same source as the training data (Flickr).
- The 70+ group on the full pipeline.
- Accessories and cosmetics one by one.
These gaps are listed on the method page and will be updated with each measurement.
The whole industry has these limits
In 2024, NIST evaluated six age estimators submitted by developers on about eleven million photos. Its conclusion: accuracy is strongly influenced by the algorithm, sex, image quality, region of birth and age, and no algorithm was better everywhere. That is one reason Face-Age refuses any use for age verification, hiring, insurance, credit or access control.
Sources
- Face-Age validation, third measurement, 18 October 2026: method and limits.
- Kärkkäinen and Joo, FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age, arXiv:1908.04913.
- NIST IR 8525, Face Analysis Technology Evaluation: Age Estimation and Verification, Hanaoka, Grother et al., May 2024.
- CNIL, facial recognition: detecting a face is not identifying a person.